Changelog ✽ Eidolon ✽ LAPLACE

Sep 9, 2026

Stream recall: a new broadcast catalog, stream_catalog.json (built by laplace-build-stream-catalog), paired with a recall_streams tool that answers whole-broadcast questions such as "what did you stream last week"; dates and periods are resolved in Chinese and English against PERSONA_TIMEZONE, at most twice per user turn
Stream archive now runs before persona facts in the voice pipeline, and recap questions read the catalog directly, skipping embedding, reranking, and persona-fact retrieval
laplace-search --recall --today replays the same recall path against a fixed clock
Fish Audio TTS is no longer cut off when its first audio packet is slow

Mem0 is now self-hosted: vectors live in Qdrant and the engine runs embedded in the bot process, with no dependency on Mem0's cloud service
New laplace-export-mem0 / laplace-import-mem0 dump a cloud account to JSON and load it into Qdrant, idempotently and resumably
With TURN configured, the server gathers relay candidates only, so the origin's public IP never reaches a viewer's SDP; the relay runs over TCP, and unroutable remote candidates are dropped

Cross-encoder reranking now runs through OpenRouter (cohere/rerank-4-fast) on the same OPENROUTER_API_KEY and is on by default; RERANK_ENABLED=false turns it off, and RERANK_BASE_URL still points back at Cohere directly
Reranking retries immediately on a 429 and falls back quietly on failure, no longer affecting the turn's retrieval
The MCP service moved to the official SDK 2.0
Default models unified: LLM ~google/gemini-flash-latest, TTS s2.1-pro, refinement ~google/gemini-pro-latest
Pipecat's bundled demo UI is off by default; set EIDOLON_PIPECAT_DEMO_ENABLED=true to serve it at / and /client

Added a Soniox transcription glossary, soniox_context.toml, improving recognition of proper nouns and names
The offline transcription default moved to stt-async-v5
Transcription polling uses a longer HTTP timeout and tolerates transient errors

Public MCP retrieval service: a read-only Streamable HTTP endpoint at /mcp exposing two tools, search_persona_facts and search_streams, with tiered rate limits by client IP / Bilibili UID

Transcription supports automatic chunking, breaking past the Soniox per-file duration limit
laplace-ingest-facts accepts ISO date formats alongside Chinese dates
Facts ingest detects and strips Tieba floor markers (第NNN楼:) from raw fact lines

Per-viewer Mem0 management endpoints: GET /api/memories to list, POST /api/memories/forget to clear, both authenticated with loginSyncToken and isolated by Bilibili UID
Fish Audio TTS switches voices at runtime, triggered by voice commands such as "be gentler" or "be more upbeat"
LaplaceMem0MemoryService filters out tool-call messages, keeping them from polluting long-term memory

Hybrid retrieval upgrade: vector + Chinese BM25 (jieba-tokenized FTS), with optional Cohere rerank-v4.0-fast cross-encoder reranking
New laplace-eval-retrieval evaluation tool, quantifying retrieval recall@k / MRR
Opus bitrate chosen per session based on the frontend audio quality
Caching and colloquial phrasing for recent stream dates

Cloudflare Realtime TURN integration, with credentials minted on demand and rotated automatically
Viewer identity is verified with a LAPLACE Login Sync token, and signed-in users skip the anonymous rate limit
Added PERSONA_TIMEZONE to inject a timezone-aware system prompt prefix
LanceDB indexes are tracked with Git LFS and bundled into the CI image

Stream Archive: a LanceDB-backed RAG service over past stream segments
Persona Facts: a LanceDB-backed RAG service over persona facts
First offline toolchain: a full pipeline for download, transcription, refinement, and ingestion

Repository initialized

Last updated on September 10, 2026